AI agents simulate emergence of social reality of emotion

A new paper from researchers Kentaro Nomura, Yushi Tsubamoto, and Takato Horii, published on arXiv (2605.07761), introduces a computational model that simulates how emotions become socially real. Grounded in the 'theory of constructed emotion,' the model treats social reality as a community-level consensus on emotion concepts tied to interoceptive sensations (internal bodily signals) and social interaction. The key innovation is integrating symbol emergence—where agents develop shared symbols through interaction—with active inference, a framework for adaptive behavior.

The simulation involves two agents each receiving interoceptive signals. They exchange inferred symbols (representing emotional interpretations) while simultaneously updating their internal bodily control goals and personal symbol interpretations. This dynamic process, described as 'social allostasis,' allows the agents to align their internal models through mutual adaptation. The experiment tracks how the agents' prior preferences over interoceptive states and their probability distributions over symbols evolve over time.

Results confirm that the two agents' interoceptive prior preferences converge, as do their symbol probability distributions. This convergence reflects the emergence of social reality—a shared, grounded understanding of emotion that is neither imposed from outside nor purely individual. The model demonstrates that emotional concepts can arise from joint negotiation between agents, rather than being hard-coded or learned from a single environment.

For the AI community, this work offers a blueprint for building agents that develop shared emotional semantics through interaction, which could enhance human-AI collaboration, empathetic AI, and multi-agent systems. It also provides a formal bridge between psychology (constructed emotion theory) and multi-agent reinforcement learning.